ML Model Binaries for Memory-Restricted Industrial Field Devices
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for deploying machine learning models on memory-restricted industrial field devices like PLCs are inefficient due to the need for resource-intensive model interpreters and high-level programming languages, leading to non-deterministic execution times and additional latency, which are unsuitable for real-time applications.
Innovation Solution
A method involving automatic generation of source code files tailored to the target field device, transforming them into a model binary, and deploying it efficiently using a tool chain, with model parameters included as constants or separate binaries, optimizing resource usage and ensuring real-time execution.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a model interpreter module is used to execute machine learning models on field devices, then the model can be deployed, but the system resources are consumed excessively and execution time becomes non-deterministic
Solution Approach 1:
The patent extracts the machine learning model from the high-level interpreted format and transforms it into a low-level optimized format specific to the target field device. The model is compiled into a device-specific binary that can be executed directly without requiring a general-purpose interpreter, thereby eliminating the resource overhead and non-deterministic execution characteristics of interpreters.
Solution Approach 2:
The patent changes the execution parameters by transforming the model from a high-level serialized format to a low-level optimized binary format. This parameter transformation involves converting the model representation into a format that matches the target device's architecture and execution model, enabling direct execution without interpretation overhead.
2Adaptability or versatility
If high-level programming languages are used for model deployment, then flexibility is achieved, but resource usage increases and real-time guarantees are lost
Solution Approach 1:
The patent segments the model deployment process into distinct stages: model training in a high-level environment, model export to an intermediate format, and model compilation into device-specific binary. This segmentation allows flexibility during training while optimizing for execution on the target device, separating the requirements of model development from execution efficiency.
Solution Approach 2:
The patent replaces the mechanical interpretation process with a compilation process. Instead of using high-level languages that require interpretation or virtual machine execution, the model is compiled into native machine code or optimized binary format specific to the field device, eliminating the resource-intensive interpretation mechanism.
3Reliability
If manual coding and optimization is performed for field devices, then real-time requirements are met, but development complexity and time increase significantly
Solution Approach 1:
The patent implements self-service by enabling automatic model compilation and optimization. The system automatically generates device-specific code from the high-level model representation, eliminating the need for manual coding while ensuring real-time performance requirements are met. The compilation process handles optimization automatically based on target device characteristics.
Solution Approach 2:
The patent performs preliminary actions by pre-compiling and optimizing the model during the development phase before deployment to the field device. The model is compiled into optimized binary format in advance, so that when deployed to the resource-constrained field device, it executes efficiently without requiring manual optimization, thereby reducing development complexity.
4Adaptability or versatility
If generic model interpreters are used, then various models can be executed, but memory requirements exceed available field device memory
Solution Approach 1:
The patent applies local quality by generating device-specific optimized code tailored to the target field device's architecture and memory constraints. Instead of using a generic interpreter that must support multiple model formats and device types, the system generates specialized execution code for each target device, optimizing memory usage for that specific platform while maintaining model functionality.
Data Source
AI summary
Provided is a method for executing a machine learning, model on a field device including executing basic operations of the machine learning, model divided into operation groups of basic operations according to a schedule, wherein basic operations of an operation group are executed while model parameters of a subsequent operation group are loaded.